Faithfulness and Stability of Black Box Explanations for Tabular and Image Classifiers

Local explanation rankings can depend on the classifier, masking operator and computational budget. This study compares LIME, KernelSHAP and Ablation or Occlusion in two controlled cases: Breast Cancer Wisconsin Diagnostic with a random forest and multilayer perceptron and Fashion-MNIST with a multilayer perceptron and convolutional neural network. Evaluation uses 104 shared correctly classified tabular cases and 100 class-balanced shared correctly classified images. Sensitivity analyses use 20 cases per modality, ten explanation seeds, three Gaussian noise scales, three prediction-preservation thresholds and three replacement baselines. Paired tests apply Holm adjustment across 60 method contrasts. KernelSHAP has the lowest mean tabular deletion AUC. Occlusion leads for the image multilayer perceptron and LIME for the convolutional network. Deterministic removal methods repeat exactly for nondegenerate vectors, while LIME generally has greater signed rank stability under retained input changes. Alternative image backgrounds change attribution vectors and can reverse deletion rankings. Evaluated-input costs range from 31 for tabular Ablation to approximately 15,001 for tabular KernelSHAP. These results support separate reporting of faithfulness, stability and cost under documented protocols. They establish neither a universal method ranking nor clinical utility or generalization across datasets and training seeds.